CONTENTS

DESCRIPTION

Security is best implemented through a layered onion approach. In a nutshell, what you want to do is to create as many layers of security as are convenient and then carefully monitor the system for intrusions. You do not want to overbuild your security or you will interfere with the detection side, and detection is one of the single most important aspects of any security mechanism. For example, it makes little sense to set the schg flags (see chflags(1)) on every system binary because while this may temporarily protect the binaries, it prevents an attacker who has broken in from making an easily detectable change that may result in your security mechanisms not detecting the attacker at all.

System security also pertains to dealing with various forms of attacks, including attacks that attempt to crash or otherwise make a system unusable but do not attempt to break root. Security concerns can be split up into several categories:

1. Denial of Service attacks (DoS)
2. User account compromises
3. Root compromise through accessible servers
4. Root compromise via user accounts
5. Backdoor creation

A denial of service attack is an action that deprives the machine of needed resources. Typically, DoS attacks are brute-force mechanisms that attempt to crash or otherwise make a machine unusable by overwhelming its servers or network stack. Some DoS attacks try to take advantages of bugs in the networking stack to crash a machine with a single packet. The latter can only be fixed by applying a bug fix to the kernel. Attacks on servers can often be fixed by properly specifying options to limit the load the servers incur on the system under adverse conditions. Brute-force network attacks are harder to deal with. A spoofed-packet attack, for example, is nearly impossible to stop short of cutting your system off from the Internet. It may not be able to take your machine down, but it can fill up Internet pipe.

A user account compromise is even more common than a DoS attack. Many sysadmins still run standard telnetd(8), rlogind(8), rshd(8), and ftpd(8) servers on their machines. These servers, by default, do not operate over encrypted connections. The result is that if you have any moderate-sized user base, one or more of your users logging into your system from a remote location (which is the most common and convenient way to log in to a system) will have his or her password sniffed. The attentive system administrator will analyze his remote access logs looking for suspicious source addresses even for successful logins.

One must always assume that once an attacker has access to a user account, the attacker can break root. However, the reality is that in a well secured and maintained system, access to a user account does not necessarily give the attacker access to root. The distinction is important because without access to root the attacker cannot generally hide his tracks and may, at best, be able to do nothing more than mess with the user’s files or crash the machine. User account compromises are very common because users tend not to take the precautions that sysadmins take.

System administrators must keep in mind that there are potentially many ways to break root on a machine. The attacker may know the root password, the attacker may find a bug in a root-run server and be able to break root over a network connection to that server, or the attacker may know of a bug in an SUID-root program that allows the attacker to break root once he has broken into a user’s account. If an attacker has found a way to break root on a machine, the attacker may not have a need to install a backdoor. Many of the root holes found and closed to date involve a considerable amount of work by the attacker to clean up after himself, so most attackers do install backdoors. This gives you a convenient way to detect the attacker. Making it impossible for an attacker to install a backdoor may actually be detrimental to your security because it will not close off the hole the attacker used to break in in the first place.

Security remedies should always be implemented with a multi-layered "onion peel" approach and can be categorized as follows:

1. Securing root and staff accounts
2. Securing root — root-run servers and SUID/SGID binaries
3. Securing user accounts
4. Securing the password file
5. Securing the kernel core, raw devices, and file systems
6. Quick detection of inappropriate changes made to the system
7. Paranoia

SECURING THE ROOT ACCOUNT AND SECURING STAFF ACCOUNTS

Of course, as a sysadmin you have to be able to get to root, so we open up a few holes. But we make sure these holes require additional password verification to operate. One way to make root accessible is to add appropriate staff accounts to the "wheel" group (in /etc/group). The staff members placed in the wheel group are allowed to su(1) to root. You should never give staff members native wheel access by putting them in the wheel group in their password entry. Staff accounts should be placed in a "staff" group, and then added to the wheel group via the /etc/group file. Only those staff members who actually need to have root access should be placed in the wheel group. It is also possible, when using an authentication method such as Kerberos, to use Kerberos’s .k5login file in the root account to allow a ksu(1) to root without having to place anyone at all in the wheel group. This may be the better solution since the wheel mechanism still allows an intruder to break root if the intruder has gotten hold of your password file and can break into a staff account. While having the wheel mechanism is better than having nothing at all, it is not necessarily the safest option.

An indirect way to secure the root account is to secure your staff accounts by using an alternative login access method and *’ing out the crypted password for the staff accounts. This way an intruder may be able to steal the password file but will not be able to break into any staff accounts or root, even if root has a crypted password associated with it (assuming, of course, that you have limited root access to the console). Staff members get into their staff accounts through a secure login mechanism such as kerberos(8) or ssh(1) using a private/public key pair. When you use something like Kerberos you generally must secure the machines which run the Kerberos servers and your desktop workstation. When you use a public/private key pair with SSH, you must generally secure the machine you are logging in from (typically your workstation), but you can also add an additional layer of protection to the key pair by password protecting the keypair when you create it with ssh-keygen(1). Being able to *-out the passwords for staff accounts also guarantees that staff members can only log in through secure access methods that you have set up. You can thus force all staff members to use secure, encrypted connections for all their sessions which closes an important hole used by many intruders: that of sniffing the network from an unrelated, less secure machine.

The more indirect security mechanisms also assume that you are logging in from a more restrictive server to a less restrictive server. For example, if your main box is running all sorts of servers, your workstation should not be running any. In order for your workstation to be reasonably secure you should run as few servers as possible, up to and including no servers at all, and you should run a password-protected screen blanker. Of course, given physical access to a workstation, an attacker can break any sort of security you put on it. This is definitely a problem that you should consider but you should also consider the fact that the vast majority of break-ins occur remotely, over a network, from people who do not have physical access to your workstation or servers.

.Fx now defaults to running talkd(8), comsat(8), and fingerd(8) in a sandbox. Another program which may be a candidate for running in a sandbox is named(8). The default rc.conf includes the arguments necessary to run named(8) in a sandbox in a commented-out form. Depending on whether you are installing a new system or upgrading an existing system, the special user accounts used by these sandboxes may not be installed. The prudent sysadmin would research and implement sandboxes for servers whenever possible.

There are a number of other servers that typically do not run in sandboxes: sendmail(8), popper(8), imapd(8), ftpd(8), and others. There are alternatives to some of these, but installing them may require more work then you are willing to put (the convenience factor strikes again). You may have to run these servers as root and rely on other mechanisms to detect break-ins that might occur through them.

SECURING THE PASSWORD FILE

The best way to detect an incursion is to look for modified, missing, or unexpected files. The best way to look for modified files is from another (often centralized) limited-access system. Writing your security scripts on the extra-secure limited-access system makes them mostly invisible to potential attackers, and this is important. In order to take maximum advantage you generally have to give the limited-access box significant access to the other machines in the business, usually either by doing a read-only NFS export of the other machines to the limited-access box, or by setting up SSH keypairs to allow the limit-access box to SSH to the other machines. Except for its network traffic, NFS is the least visible method — allowing you to monitor the file systems on each client box virtually undetected. If your limited-access server is connected to the client boxes through a switch, the NFS method is often the better choice. If your limited-access server is connected to the client boxes through a hub or through several layers of routing, the NFS method may be too insecure (network-wise) and using SSH may be the better choice even with the audit-trail tracks that SSH lays.

Once you give a limit-access box at least read access to the client systems it is supposed to monitor, you must write scripts to do the actual monitoring. Given an NFS mount, you can write scripts out of simple system utilities such as find(1) and md5(1). It is best to physically md5(1) the client-box files boxes at least once a day, and to test control files such as those found in /etc and /usr/local/etc even more often. When mismatches are found relative to the base MD5 information the limited-access machine knows is valid, it should scream at a sysadmin to go check it out. A good security script will also check for inappropriate SUID binaries and for new or deleted files on system partitions such as / and /usr.

When using SSH rather than NFS, writing the security script is much more difficult. You essentially have to scp(1) the scripts to the client box in order to run them, making them visible, and for safety you also need to scp(1) the binaries (such as find(1)) that those scripts use. The sshd(8) daemon on the client box may already be compromised. All in all, using SSH may be necessary when running over unsecure links, but it is also a lot harder to deal with.

A good security script will also check for changes to user and staff members access configuration files: .rhosts, .shosts, .ssh/authorized_keys and so forth, files that might fall outside the purview of the MD5 check.

If you have a huge amount of user disk space it may take too long to run through every file on those partitions. In this case, setting mount flags to disallow SUID binaries on those partitions is a good idea. The nosuid option (see mount(8)) is what you want to look into. I would scan them anyway at least once a week, since the object of this layer is to detect a break-in whether or not the break-in is effective.

Process accounting (see accton(8)) is a relatively low-overhead feature of the operating system which I recommend using as a post-break-in evaluation mechanism. It is especially useful in tracking down how an intruder has actually broken into a system, assuming the file is still intact after the break-in occurs.

SPECIAL SECTION ON DoS ATTACKS

1. Limiting server forks
2. Limiting springboard attacks (ICMP response attacks, ping broadcast, etc.)
3. Kernel Route Cache

A common DoS attack is against a forking server that attempts to cause the server to eat processes, file descriptors, and memory until the machine dies. The inetd(8) server has several options to limit this sort of attack. It should be noted that while it is possible to prevent a machine from going down it is not generally possible to prevent a service from being disrupted by the attack. Read the inetd(8) manual page carefully and pay specific attention to the -c -, -C , and -R options. Note that spoofed-IP attacks will circumvent the -C option to inetd(8), so typically a combination of options must be used. Some standalone servers have self-fork-limitation parameters.

The sendmail(8) daemon has its -OMaxDaemonChildren option which tends to work much better than trying to use sendmail 8 ’s load limiting options due to the load lag. You should specify a MaxDaemonChildren parameter when you start sendmail(8) high enough to handle your expected load but not so high that the computer cannot handle that number of sendmail ’s without falling on its face. It is also prudent to run sendmail(8) in "queued" mode (-ODeliveryMode=queued) and to run the daemon ("sendmail-bd") separate from the queue-runs ("sendmail-q15m"). If you still want real-time delivery you can run the queue at a much lower interval, such as -q1m , but be sure to specify a reasonable MaxDaemonChildren option for that sendmail(8) to prevent cascade failures.

The syslogd(8) daemon can be attacked directly and it is strongly recommended that you use the -s option whenever possible, and the -a option otherwise.

You should also be fairly careful with connect-back services such as tcpwrapper’s reverse-identd, which can be attacked directly. You generally do not want to use the reverse-ident feature of tcpwrappers for this reason.

It is a very good idea to protect internal services from external access by firewalling them off at your border routers. The idea here is to prevent saturation attacks from outside your LAN, not so much to protect internal services from network-based root compromise. Always configure an exclusive firewall, i.e., ' firewall everything except ports A, B, C, D, and M-Z '. This way you can firewall off all of your low ports except for certain specific services such as named(8) (if you are primary for a zone), talkd(8), sendmail(8), and other internet-accessible services. If you try to configure the firewall the other way — as an inclusive or permissive firewall, there is a good chance that you will forget to "close" a couple of services or that you will add a new internal service and forget to update the firewall. You can still open up the high-numbered port range on the firewall to allow permissive-like operation without compromising your low ports. Also take note that
.Fx allows you to control the range of port numbers used for dynamic binding via the various net.inet.ip.portrange sysctl’s (""sysctlnet.inet.ip.portrange""), which can also ease the complexity of your firewall’s configuration. I usually use a normal first/last range of 4000 to 5000, and a hiport range of 49152 to 65535, then block everything under 4000 off in my firewall (except for certain specific internet-accessible ports, of course).

Another common DoS attack is called a springboard attack — to attack a server in a manner that causes the server to generate responses which then overload the server, the local network, or some other machine. The most common attack of this nature is the ICMP PING BROADCAST attack. The attacker spoofs ping packets sent to your LAN’s broadcast address with the source IP address set to the actual machine they wish to attack. If your border routers are not configured to stomp on ping’s to broadcast addresses, your LAN winds up generating sufficient responses to the spoofed source address to saturate the victim, especially when the attacker uses the same trick on several dozen broadcast addresses over several dozen different networks at once. Broadcast attacks of over a hundred and twenty megabits have been measured. A second common springboard attack is against the ICMP error reporting system. By constructing packets that generate ICMP error responses, an attacker can saturate a server’s incoming network and cause the server to saturate its outgoing network with ICMP responses. This type of attack can also crash the server by running it out of
.Vt mbuf Ns ’s , especially if the server cannot drain the ICMP responses it generates fast enough. The
.Fx kernel has a new kernel compile option called ICMP_BANDLIM which limits the effectiveness of these sorts of attacks. The last major class of springboard attacks is related to certain internal inetd(8) services such as the UDP echo service. An attacker simply spoofs a UDP packet with the source address being server A’s echo port, and the destination address being server B’s echo port, where server A and B are both on your LAN. The two servers then bounce this one packet back and forth between each other. The attacker can overload both servers and their LANs simply by injecting a few packets in this manner. Similar problems exist with the internal chargen port. A competent sysadmin will turn off all of these inetd 8 -internal test services.

HISTORY